Traffic Prediction Analyses And Intelligent Base Station Management In Greener Cellular Networks

Cellular networks are facing the challenges to survive from huge energy wastage, while meet-ing the explosive traffic growth. The energy issue is mainly due to the fact that power consump-tion of one base station (BS) does not merely depend on the traffic loads within its coverage, and because of the existence of auxiliary devices, BSs still need to consume significant power even when traffic loads become low. Therefore, one of the most promsing research methods towards greener cellular networks is to design traffic-aware cellular networks, which adapt the working sta-tus of BS to traffic patterns. In this disseration, we focus on this appealing topic, and envision a softward-defined cellular network architecture. Within this architecture, we carry out some specific researches, spanning from traffic characteristic analyses, traffic prediction algorithms, to intelligent BS management strategies. The contents are listed as follows:We firstly aim to re-examine the characteristics of cellular network traffic, based on practical traffic "big data". In this regard, we exploit entropy theory to analyze the predictability of voice, text and data traffic in cellular networks, and validate the possiblity to predict the traffic using temporal, spatial, and inter-voice/text information. Afterwards, we study the statistical patterns of instaneous messaging service in mobile Internet, and demonstrate its heavy-tailness. We also testify the accuracy to use a-stable models to describe the service-level traffic within each BS's coverage. At last of this part, we show the sparsity embeded in cellular network traffic.In the second part, we take advantage of the aforementioned analysis results of traffic char-acteristics, and investigate practical traffic prediction algorithms. We firstly give a tempo-spatial compressive sensing based traffic prediction algorithm, and validate its superior performance to predict voice and text traffic. As for the data traffic, we reveal that the randomness of prediction error could be approximated by Gaussian noise, when applying a-stable model based linear predic-tion algorithms. We then develop an alternative direction method, which is embeded with a-stable model based linear prediction algorithm and dictionary learning algorithm, and takes advantage of the charateristics unveiled before. Besides, we validate its prediction accuracy improvement and the robustness of the proposed framework through extensive simulation results.In the third part, we design two intelligent traffic-aware BS management schemes. In the case where traffic information are meticously predicted using previous algorithms, we propose a grid-based energy saving scheme to improve the energy efficiency through turning some BSs into sleeping mode. Simulation results with real traffic load statistics manifest large improvement of energy efficiency. In the other case where traffic information is unknown a prior, we formulate the BS switching operation scheme under traffic variations as a Markov decision process, and de-sign a reinforcement learning framework based energy saving scheme. Furthermore, to speed up the ongoing learning process, a transfer actor-critic algorithm (TACT), which utilizes the trans-ferred learning expertise in historical periods or neighboring regions, is proposed and provably converges. In the end, extensive simulations under various practical configurations show that the proposed TACT algorithm contributes to a performance jumpstart and demonstrates the feasibility of significant energy efficiency improvement at the expense of tolerable delay performance.